Encoder-decoder Model For Multi-aspect Sentiment Classification

In recent years,the rapid development of informatization in various fields,especially the coverage of social networks and the popularity of online consumption,the number of online comments has increased explosively.Compared with traditional word-of-mouth transmission,online comments have better timeliness and universality.There,referring information of online comments has become an important means to fully recognize things and rationalize consumption.Since users who post or read comment will pay attention to the object in the comment from many different aspects,generating sentiment label towards the whole text or a single aspect of the object can no longer meet the need for fine-grained mining of the information expressed by the comments.Therefore,the task of multi-aspect sentiment classification for comments has become an important research topic with practical needs.The goal of multi-aspect sentiment classification is to generate corresponding sentiment label for each given aspect.In this thesis,based on the Encoder-Decoder model,the multi-aspect sentiment classification is converted into a task which maps text sequence to sentiment sequence.At the same time,in order to solve the problem of correspondence between sentiment label and aspect,this thesis proposes a target embedding method based on 1D-Convolution aspect expression.This method uses a convolutional network to extract feature from different words that express a certain aspect,and embeds the feature into the unified semantic vector generated by the Encoder.The embedding result is used as the input of the time step corresponding to the aspect to generate the sentiment label.The attention mechanism has achieved great success in the field of natural language processing.For a specific target,extracting the corresponding semantics from the text can filter out the noise caused by irrelevant information.In this thesis,an attention filtering layer is designed,based on the characteristics that the length of the aspect sequence is fixed.Different from Bahdua attention,the structure only uses output sequence of Encoder and aspect expression sequence to caculate the attention weight,and generate a semantic sequence corresponding to the aspect sequence according to the attention weight matrix.The sequence is used as the input of the Decoder to generate a sentiment label sequence.In this thesis,a new approach based on the Encoder-Decoder model and two improvements according to the characteristics of the task are proposed for multi-aspect sentiment classification.Experiments show that these methods can effectively improve performance on this task.It can also provide support for cement analysis and statistics,and also provide references for the resolution of other similar tasks.